feat!: return flash attention varlen auxiliary outputs

[voltjia]

Summary

• Keep the basic flash_attn_varlen_func(..., out) overload unchanged and out-only.
• Extend only the full interface with trailing optional softmax_lse and s_dmask outputs.
• Require both auxiliary outputs exactly when return_attn_probs=true, and populate them in the NVIDIA Torch implementation.

Motivation

return_attn_probs was already public, but the full interface omitted FlashAttention's corresponding auxiliary outputs. The basic overload should continue to model the default out-only call.

API Alignment

InfiniOps API	Alignment	Reference
flash_attn_varlen_func(..., out)	FlashAttention default call returns only out	FlashAttention source
flash_attn_varlen_func(..., return_attn_probs, ..., out, softmax_lse?, s_dmask?)	FlashAttention returns (out, softmax_lse, S_dmask) only when return_attn_probs=True	FlashAttention source

Type of Change

• feat
• fix
• perf
• refactor
• test
• docs
• build / ci
• chore
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

ssh nvidia, accelerator-dev/nvidia:latest
NVIDIA + Torch targeted integration wheel: passed
Wheel SHA-256: f72af45d2439d0c2459ab08da7461e9635f2c434e22b0ee0cce03fe0c6ff353d
tests/test_generate_wrappers.py + tests/test_generate_torch_ops.py: 35 passed
tests/test_flash_attn_varlen_func.py: 20 passed
clang-format 21.1.8 and Ruff 0.15.22: passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	targeted NVIDIA + Torch wheel build passed	20 passed
Iluvatar	No	N/A	N/A
MetaX	No	N/A	N/A
Cambricon	No	N/A	N/A
Moore	No	N/A	N/A
Ascend	No	N/A	N/A

Benchmark / Performance Impact

N/A.

Notes for Reviewers

The basic overload remains out-only. The full overload uses std::optional<Tensor> only for the conditional auxiliary outputs.

This PR is based directly on master. It contains no Moore files and has no dependency on #819; a future Moore implementation can adapt to the public interface independently.